A 1.27 GB Qwen that says yes more often: Unlocked GGUF shows the coding cost

A tiny local Qwen build with fewer refusals and less “Wait… no…” is the promise of Qwen3.5-2B-Unlocked GGUF. Community author angronain has published weight edits to Qwen3.5-2B, with a recommended 1.27 GB Q4_K_M download and instructions for Ollama and llama.cpp.

The repository was created on November 3 at 20:21 UTC. Its English card says the changes use directional ablations directly in the weights, without fine-tuning or training data: one targets refusal behavior, while another damps selected self-doubt expressions in thinking traces. Those are the author’s descriptions of the method and intended effect.

Small enough to try locally

The package includes larger quantizations, an Ollama Modelfile and a separate optional vision projector. The documented Ollama route downloads the files, creates a local model and disables thinking for ordinary chat. The llama.cpp example uses the repository’s Q4_K_M build, its chat template and a presence penalty of 1.5.

The author recommends non-thinking mode for chat and a reasoning budget when thinking is enabled. There is no need to interpret the “Unlocked” name as improved general intelligence: the card describes a targeted behavioral edit with measurable costs in its own tests.

More compliance comes with tradeoffs

On ten executed coding tasks averaged over four runs at Q6_K, the author reports a fall from 5.0 successful tasks for the original model to 3.75 for the edited version. Unfinished thinking runs fall from 16 of 22 to 10 of 22 at an 8,000-token cap, leaving substantial failures.

The refusal table reports zero refusals on 200 held-out prompts for Unlocked, but compares that with 46 refusals on a different 48-prompt original-model test set. Those unequal cohorts do not establish a controlled percentage improvement. AITrending has not reproduced these evaluations.

The author also discloses factual slips, basic vision limitations and cases where retrieval prompts still produce invented answers. The practical recommendation is therefore specific: examine this compact build if its refusal and thinking behavior fits your local workflow, while keeping the disclosed coding and reliability costs in view.